What happened
In Arize Phoenix (AI observability & evaluation platform) from 13.0.0 to <15.7.0, the applyChatTemplate GraphQL query passes attacker-controlled F-string templates to FStringTemplateFormatter._format while DotDict.__getattr__ fails to restrict traversal through Python private attributes. A crafted template can traverse into objects and evaluate arbitrary code, giving unauthenticated/remote code execution. CVSS 10.0 (Critical). Fixed in 15.7.0.
Why it matters
Phoenix is widely used to trace, evaluate and observe LLM/agent pipelines, and it exposes a GraphQL surface. An attacker who can reach the applyChatTemplate query and inject an F-string template achieves code execution inside the observability platform, which sits adjacent to model logs, evals, prompts and secrets. Critical severity in a core AI-operations component makes this Tier A.
Attack vector
Unauthenticated/remote call to the applyChatTemplate GraphQL query with a crafted F-string template that traverses DotDict private attributes to evaluate code server-side
Affected systems
Arize Phoenix 13.0.0 to <15.7.0
Mitigation
Upgrade to Arize Phoenix 15.7.0 or later; restrict access to the GraphQL endpoint